🌱 Force Cache Refresh when Server was not found by name.

hcloud/cloud_test.go

@@ -34,6 +34,7 @@ import (
 	"github.com/syself/hetzner-cloud-controller-manager/internal/annotation"
 	"github.com/syself/hetzner-cloud-controller-manager/internal/credentials"
 	"github.com/syself/hetzner-cloud-controller-manager/internal/hcops"
+	robotclient "github.com/syself/hetzner-cloud-controller-manager/internal/robot/client"
 	hrobot "github.com/syself/hrobot-go"
 	"github.com/syself/hrobot-go/models"
 	corev1 "k8s.io/api/core/v1"
@@ -44,7 +45,7 @@ type testEnv struct {
 	Server      *httptest.Server
 	Mux         *http.ServeMux
 	Client      *hcloud.Client
-	RobotClient hrobot.RobotClient
+	RobotClient robotclient.Client
 }
 
 func (env *testEnv) Teardown() {
@@ -70,7 +71,7 @@ func newTestEnv() testEnv {
 		Server:      server,
 		Mux:         mux,
 		Client:      client,
-		RobotClient: robotClient,
+		RobotClient: robotclient.New(robotClient),
 	}
 }
 

hcloud/instances_test.go

@@ -26,6 +26,7 @@ import (
 
 	"github.com/hetznercloud/hcloud-go/v2/hcloud"
 	"github.com/hetznercloud/hcloud-go/v2/hcloud/schema"
+	"github.com/syself/hetzner-cloud-controller-manager/internal/robot/client/cache"
 	"github.com/syself/hrobot-go/models"
 	corev1 "k8s.io/api/core/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
@@ -173,6 +174,128 @@ func TestInstances_InstanceExists(t *testing.T) {
 	}
 }
 
+func TestInstances_InstanceExistsRobotServerCreatedAfterCacheFill(t *testing.T) {
+	env := newTestEnv()
+	defer env.Teardown()
+
+	resetEnv := Setenv(t,
+		"ROBOT_USER_NAME", "user",
+		"ROBOT_PASSWORD", "pass",
+		"CACHE_TIMEOUT", "1h",
+	)
+	defer resetEnv()
+
+	// servers backs the Robot list response and is mutated during the test.
+	servers := make([]models.Server, 0, 2)
+	servers = append(servers, models.Server{
+		ServerIP:      "123.123.123.123",
+		ServerIPv6Net: "2a01:f48:111:4221::",
+		ServerNumber:  321,
+		Name:          "bm-existing",
+	})
+	env.Mux.HandleFunc("/robot/server", func(w http.ResponseWriter, _ *http.Request) {
+		responses := make([]models.ServerResponse, 0, len(servers))
+		for _, server := range servers {
+			responses = append(responses, models.ServerResponse{Server: server})
+		}
+		json.NewEncoder(w).Encode(responses)
+	})
+
+	robotClient, err := cache.NewCachedRobotClient(t.TempDir(), env.Server.Client(), env.Server.URL+"/robot")
+	if err != nil {
+		t.Fatalf("Unexpected error creating cached robot client: %v", err)
+	}
+
+	instances := newInstances(env.Client, robotClient, AddressFamilyIPv4, 0)
+
+	// Warm the cache while bm-new does not exist yet.
+	exists, err := instances.InstanceExists(context.TODO(), &corev1.Node{
+		ObjectMeta: metav1.ObjectMeta{Name: "bm-existing"},
+	})
+	if err != nil {
+		t.Fatalf("Unexpected error warming cache: %v", err)
+	}
+	if !exists {
+		t.Fatal("Expected bm-existing to exist")
+	}
+
+	servers = append(servers, models.Server{
+		ServerIP:      "123.123.123.124",
+		ServerIPv6Net: "2a01:f48:111:4222::",
+		ServerNumber:  322,
+		Name:          "bm-new",
+	})
+
+	exists, err = instances.InstanceExists(context.TODO(), &corev1.Node{
+		ObjectMeta: metav1.ObjectMeta{Name: "bm-new"},
+	})
+	if err != nil {
+		t.Fatalf("Unexpected error for bm-new: %v", err)
+	}
+	if !exists {
+		t.Fatal("Expected bm-new to exist after it was created")
+	}
+}
+
+func TestInstances_InstanceExistsRobotServerRepeatedMissingNameSkipsForceRefresh(t *testing.T) {
+	env := newTestEnv()
+	defer env.Teardown()
+
+	resetEnv := Setenv(t,
+		"ROBOT_USER_NAME", "user",
+		"ROBOT_PASSWORD", "pass",
+		"CACHE_TIMEOUT", "1h",
+	)
+	defer resetEnv()
+
+	robotServerListCalls := 0
+	env.Mux.HandleFunc("/robot/server", func(w http.ResponseWriter, _ *http.Request) {
+		robotServerListCalls++
+		json.NewEncoder(w).Encode([]models.ServerResponse{
+			{
+				Server: models.Server{
+					ServerIP:      "123.123.123.123",
+					ServerIPv6Net: "2a01:f48:111:4221::",
+					ServerNumber:  321,
+					Name:          "bm-existing",
+				},
+			},
+		})
+	})
+
+	robotClient, err := cache.NewCachedRobotClient(t.TempDir(), env.Server.Client(), env.Server.URL+"/robot")
+	if err != nil {
+		t.Fatalf("Unexpected error creating cached robot client: %v", err)
+	}
+
+	instances := newInstances(env.Client, robotClient, AddressFamilyIPv4, 0)
+	node := &corev1.Node{
+		ObjectMeta: metav1.ObjectMeta{Name: "bm-missing"},
+	}
+
+	exists, err := instances.InstanceExists(context.TODO(), node)
+	if err != nil {
+		t.Fatalf("Unexpected error on first miss: %v", err)
+	}
+	if exists {
+		t.Fatal("Expected bm-missing to be absent on first lookup")
+	}
+	if robotServerListCalls != 2 {
+		t.Fatalf("Expected 2 Robot list calls after first miss, got %d", robotServerListCalls)
+	}
+
+	exists, err = instances.InstanceExists(context.TODO(), node)
+	if err != nil {
+		t.Fatalf("Unexpected error on second miss: %v", err)
+	}
+	if exists {
+		t.Fatal("Expected bm-missing to be absent on second lookup")
+	}
+	if robotServerListCalls != 2 {
+		t.Fatalf("Expected repeated miss to skip force refresh, got %d Robot list calls", robotServerListCalls)
+	}
+}
+
 func TestInstances_InstanceShutdown(t *testing.T) {
 	env := newTestEnv()
 	defer env.Teardown()

hcloud/util.go

@@ -73,11 +73,29 @@ func getRobotServerByName(c robotclient.Client, node *corev1.Node) (server *mode
 
 	for i, s := range serverList {
 		if s.Name == node.Name {
-			server = &serverList[i]
+			return &serverList[i], nil
 		}
 	}
 
-	return server, nil
+	// CAPH changes the Robot server name during provisioning. This means the cache of the
+	// server-name-to-server-ID mapping could be outdated. Force one uncached Robot list reload for
+	// that name to handle the rename. The Robot client suppresses repeated forced refreshes for the
+	// same missing name until the normal Robot list cache timeout has elapsed.
+	serverList, err = c.ServerGetListForceRefresh(string(node.Name))
+	if err != nil {
+		hcops.HandleRateLimitExceededError(err, node)
+		return nil, fmt.Errorf("%s: force refresh after cache miss: %w", op, err)
+	}
+
+	for i, s := range serverList {
+		if s.Name == node.Name {
+			// Server was found after cache refresh
+			return &serverList[i], nil
+		}
+	}
+
+	// No server found.
+	return nil, nil
 }
 
 func getRobotServerByID(c robotclient.Client, id int, node *corev1.Node) (s *models.Server, e error) {

internal/mocks/robot.go

@@ -14,6 +14,10 @@ func (m *RobotClient) ServerGetList() ([]models.Server, error) {
 	return getRobotServers(args, 0), args.Error(1)
 }
 
+func (m *RobotClient) ServerGetListForceRefresh(_ string) ([]models.Server, error) {
+	return m.ServerGetList()
+}
+
 func (m *RobotClient) SetCredentials(_, _ string) error {
 	args := m.Called()
 	return args.Error(3)

internal/robot/client/adapter.go

@@ -0,0 +1,26 @@
+package client
+
+import (
+	hrobot "github.com/syself/hrobot-go"
+	"github.com/syself/hrobot-go/models"
+)
+
+var _ Client = &adapter{}
+
+type adapter struct {
+	hrobot.RobotClient
+}
+
+// New wraps a plain Robot client so it satisfies the local Client interface.
+func New(robotClient hrobot.RobotClient) Client {
+	if robotClient == nil {
+		return nil
+	}
+	return &adapter{RobotClient: robotClient}
+}
+
+// ServerGetListForceRefresh falls back to the plain list call because the
+// uncached behavior only exists in the cache-backed client implementation.
+func (a *adapter) ServerGetListForceRefresh(_ string) ([]models.Server, error) {
+	return a.ServerGetList()
+}

internal/robot/client/cache/client.go

@@ -18,6 +18,7 @@ const (
 	robotUserNameENVVar = "ROBOT_USER_NAME"
 	robotPasswordENVVar = "ROBOT_PASSWORD"
 	cacheTimeoutENVVar  = "CACHE_TIMEOUT"
+	defaultCacheTimeout = 5 * time.Minute
 )
 
 var _ robotclient.Client = &cacheRobotClient{}
@@ -27,10 +28,16 @@ type cacheRobotClient struct {
 	timeout     time.Duration
 
 	lastUpdate time.Time
+	now        func() time.Time
 
 	// cache
 	l []models.Server
 	m map[int]*models.Server
+
+	// forcedRefreshServerNames stores when a node name last triggered a forced Robot server list
+	// refresh. While that timestamp is still within the cache timeout window, repeated lookups for
+	// the same missing server name skip the extra uncached Robot API call.
+	forcedRefreshServerNames map[string]time.Time
 }
 
 // NewCachedRobotClient creates a new robot client with caching enabled.
@@ -51,7 +58,7 @@ func NewCachedRobotClient(rootDir string, httpClient *http.Client, baseURL strin
 	}
 
 	if cacheTimeout == 0 {
-		cacheTimeout = 5 * time.Minute
+		cacheTimeout = defaultCacheTimeout
 	}
 
 	credentialsDir := credentials.GetDirectory(rootDir)
@@ -85,6 +92,7 @@ func NewCachedRobotClient(rootDir string, httpClient *http.Client, baseURL strin
 	handler := &cacheRobotClient{}
 	handler.timeout = cacheTimeout
 	handler.robotClient = c
+	handler.now = time.Now
 	return handler, nil
 }
 
@@ -105,7 +113,7 @@ func (c *cacheRobotClient) ServerGet(id int) (*models.Server, error) {
 		}
 
 		// set time of last update
-		c.lastUpdate = time.Now()
+		c.lastUpdate = c.currentTime()
 	}
 
 	server, found := c.m[id]
@@ -134,30 +142,91 @@ func (c *cacheRobotClient) ServerGetList() ([]models.Server, error) {
 		}
 
 		// set time of last update
-		c.lastUpdate = time.Now()
+		c.lastUpdate = c.currentTime()
 	}
 
 	return c.l, nil
 }
 
+// ServerGetListForceRefresh invalidates the current cache and reloads the list
+// from Robot unless nodeName already triggered a forced refresh within the
+// current timeout window.
+func (c *cacheRobotClient) ServerGetListForceRefresh(nodeName string) ([]models.Server, error) {
+	if nodeName != "" && c.nodeHasAlreadyForcedRefresh(nodeName) {
+		return c.ServerGetList()
+	}
+
+	c.m = nil
+	list, err := c.ServerGetList()
+	if err != nil {
+		return nil, err
+	}
+	if nodeName != "" {
+		c.nodeTriggeredForcedRefresh(nodeName)
+	}
+	return list, nil
+}
+
+// nodeHasAlreadyForcedRefresh reports whether nodeName already triggered a
+// forced refresh within the current cache timeout window.
+func (c *cacheRobotClient) nodeHasAlreadyForcedRefresh(nodeName string) bool {
+	if c.forcedRefreshServerNames == nil {
+		return false
+	}
+
+	forcedAt, found := c.forcedRefreshServerNames[nodeName]
+	if !found {
+		return false
+	}
+	if c.currentTime().After(forcedAt.Add(c.forceRefreshTimeout())) {
+		delete(c.forcedRefreshServerNames, nodeName)
+		return false
+	}
+	return true
+}
+
+// nodeTriggeredForcedRefresh records that nodeName already triggered a forced
+// refresh.
+func (c *cacheRobotClient) nodeTriggeredForcedRefresh(nodeName string) {
+	if c.forcedRefreshServerNames == nil {
+		c.forcedRefreshServerNames = make(map[string]time.Time)
+	}
+	c.forcedRefreshServerNames[nodeName] = c.currentTime()
+}
+
 func (c *cacheRobotClient) shouldSync() bool {
 	// map is nil means we have no cached value yet
 	if c.m == nil {
 		c.m = make(map[int]*models.Server)
 		return true
 	}
-	if time.Now().After(c.lastUpdate.Add(c.timeout)) {
+	if c.currentTime().After(c.lastUpdate.Add(c.timeout)) {
 		return true
 	}
 	return false
 }
 
+func (c *cacheRobotClient) currentTime() time.Time {
+	if c.now == nil {
+		return time.Now()
+	}
+	return c.now()
+}
+
+func (c *cacheRobotClient) forceRefreshTimeout() time.Duration {
+	if c.timeout == 0 {
+		return defaultCacheTimeout
+	}
+	return c.timeout
+}
+
 func (c *cacheRobotClient) SetCredentials(username, password string) error {
 	err := c.robotClient.SetCredentials(username, password)
 	if err != nil {
 		return err
 	}
 	// The credentials have been updated, so we need to invalidate the cache.
 	c.m = nil
+	c.forcedRefreshServerNames = nil
 	return nil
 }